I
ndo
ne
s
i
a
n J
o
ur
na
l
o
f
E
l
e
c
t
r
i
c
a
l
E
ng
i
ne
e
r
i
ng
a
nd C
o
m
p
u
t
e
r
Sc
i
e
nc
e
V
o
l.
11
,
N
o.
2
,
A
ug
us
t
201
8
, p
p
.
55
8
~
56
6
I
S
S
N
:
2502
-
4752,
D
O
I
:
10.
11
591/
i
j
e
e
c
s
.
v
11.
i
2.
pp
558
-
5
66
558
Jou
r
n
al
h
om
e
p
age
:
h
ttp
:
//ia
e
s
c
or
e
.
c
om
/
j
our
nal
s
/
i
nde
x
.
php/
i
j
eecs
Ca
n Co
nv
o
lut
io
n
Neura
l Ne
t
w
o
rk
(
CNN)
T
riu
m
ph
i
n E
a
r
Reco
g
nit
io
n o
f
U
nif
o
r
m
I
llu
m
ina
t
i
o
n I
nv
a
ria
nt
?
N
u
r
s
u
r
ia
t
i J
a
m
il
1
,
A
li A
b
d
A
l
m
is
r
e
b
2
,
Sy
e
d M
o
hd
Z
a
hi
d Sy
e
d
Z
a
i
na
l
A
r
i
f
f
i
n
3
,
N.
M
d
Di
n
4
,
R
a
s
eed
a
H
a
m
za
h
5
1
,3
,5
F
acu
l
t
y
o
f
C
o
m
p
u
t
er
an
d
M
at
h
em
at
i
c
al
S
ci
en
ces
,
U
n
iv
e
r
s
iti T
e
k
n
o
lo
g
i M
A
R
A
,
40
45
0 S
ha
h A
l
a
m
,
S
e
l
a
ng
or
,
M
a
l
a
y
s
i
a
2
,
4
C
o
l
l
eg
e o
f
G
r
ad
u
at
e S
t
u
d
i
es
,
U
n
i
v
er
s
i
t
i
T
en
ag
a N
a
s
i
o
n
al
,
J
al
an
I
K
R
A
M
-
UNI
T
E
N,
4
3
0
0
0
Ka
j
a
n
g
,
M
a
l
a
y
si
a
A
rt
i
cl
e I
n
f
o
AB
S
T
RAC
T
A
r
tic
le
h
is
to
r
y
:
R
ecei
v
ed
M
ar
1,
2018
R
e
v
i
s
e
d
A
pr
21
,
201
8
Ac
cep
t
ed
M
a
y
1
,
2018
C
ur
r
e
nt
de
e
p c
o
nv
ol
ut
i
o
n ne
ur
a
l
ne
t
w
or
k
(
C
N
N
)
ha
s
s
how
n t
o a
c
hi
e
v
e
s
upe
r
i
or
pe
r
f
or
m
a
nc
e
on a
num
be
r
of
c
o
m
put
e
r
v
i
s
i
on t
a
s
k
s
s
uc
h a
s
i
m
a
ge
r
e
c
o
g
n
itio
n
,
c
la
s
s
if
ic
a
tio
n
a
n
d
o
b
je
c
t d
e
te
c
tio
n
.
T
h
e d
eep
n
et
w
o
r
k
w
as
al
s
o
t
es
t
ed
f
o
r
v
i
ew
-
i
nva
r
i
a
nc
e
,
r
obus
t
ne
s
s
a
nd i
l
l
um
i
na
t
i
o
n i
nv
a
r
i
a
nc
e
.
H
ow
e
v
e
r
,
t
h
e C
N
N
ar
ch
i
t
ect
u
r
e
h
as
t
h
u
s
f
ar
o
n
l
y
b
een
t
es
t
ed
o
n
n
o
n
-
u
ni
f
or
m
i
l
l
um
i
na
t
i
on i
nv
a
r
i
a
nt
.
C
a
n C
N
N
pe
r
f
or
m
e
qua
l
l
y
w
e
l
l
f
or
v
e
r
y
und
e
r
e
x
pos
e
d
or
ov
e
r
e
x
pos
e
d i
m
a
g
e
s
or
k
now
n a
s
u
ni
f
or
m
i
l
l
um
i
na
t
i
on i
nv
a
r
ia
n
t?
T
h
is
is
t
h
e g
ap
t
h
at
w
e ar
e ad
d
r
es
s
i
n
g
i
n
t
h
i
s
p
ap
er
.
I
n
o
u
r
w
o
r
k
,
w
e
co
l
l
ect
ed
ear
i
m
a
g
e
s
unde
r
di
f
f
e
r
e
nt
uni
f
or
m
i
l
l
um
i
na
t
i
on c
o
ndi
t
i
o
ns
w
i
t
h l
um
e
ns
or
l
ux
v
a
l
ue
s
r
a
ng
i
ng
f
r
o
m
2 l
ux
t
o 10,
700 l
ux
.
A
t
ot
a
l
of
1,
100 l
e
f
t
a
nd r
i
g
ht
e
a
r
i
ma
g
e
s
f
r
o
m 5
5
s
u
bj
e
c
t
s
a
r
e
c
a
pt
ur
e
d
un
de
r
na
t
ur
a
l
i
l
l
um
i
na
t
i
on
c
ond
i
t
i
o
ns
.
A
s
C
N
N
r
eq
u
i
r
es
co
n
s
i
d
er
ab
l
y
l
ar
g
e a
m
o
u
n
t
o
f
d
at
a,
t
h
e ear
i
m
ag
e
s
ar
e
f
u
r
t
h
er
r
o
t
at
ed
at
ev
er
y
5
o
an
g
l
es
t
o
g
en
er
at
e 2
5
,
3
0
0
i
m
ag
es
.
F
o
r
each
s
u
b
j
ect
,
5
0
i
m
ag
es
ar
e u
s
ed
as
v
a
l
i
d
at
i
o
n
/
t
es
t
i
n
g
d
at
a
s
e
t,
w
h
ile
th
e
r
e
m
a
in
in
g
i
m
a
g
e
s
a
r
e
us
e
d a
s
t
r
a
i
ni
ng
da
t
a
s
e
t
s
.
O
ur
pr
opos
e
d C
N
N
m
ode
l
i
s
t
he
n
t
r
a
i
ne
d f
r
om
s
c
r
a
t
c
h a
nd v
a
l
i
da
t
i
on a
nd t
e
s
t
i
ng
r
e
s
ul
t
s
s
how
e
d r
e
c
og
ni
t
i
on
accu
r
ac
y
o
f
9
7
%
.
T
h
e r
es
u
l
t
s
s
h
o
w
ed
t
h
at
1
0
0
%
accu
r
acy
i
s
ac
h
i
ev
ed
f
o
r
im
a
g
e
s
w
ith
lu
m
e
ns
r
a
ng
i
ng
a
bo
v
e
30 bu
t
ha
v
i
ng
pr
obl
e
m
w
i
t
h l
um
e
ns
l
e
s
s
t
ha
n 10 l
ux
.
Ke
y
wo
rd
s
:
C
o
n
vo
l
ut
i
o
n N
e
ur
a
l
N
e
t
w
o
r
k
E
ar
R
e
c
o
gni
t
i
o
n
U
n
i
f
or
m
I
l
l
um
i
n
a
t
i
on
I
nv
a
r
i
a
nt
L
u
me
n
s
C
opy
r
i
g
ht
©
201
8
I
ns
t
i
t
ut
e
o
f
A
d
v
anc
e
d E
ngi
ne
e
r
i
ng
an
d
Sc
i
e
nc
e
.
A
l
l
ri
g
h
t
s re
se
rv
e
d
.
Co
rre
sp
o
n
d
i
n
g
Au
t
h
o
r
:
N
u
r
s
u
r
ia
ti J
a
m
il
,
F
acu
l
t
y
o
f
C
o
m
p
u
t
er
an
d
M
at
h
e
m
a
t
i
cal
S
ci
e
n
ces
,
U
ni
ve
r
s
i
t
i
T
e
kno
l
o
gi
M
A
R
A
,
4
0
4
50
S
ha
h
A
l
a
m
,
S
e
l
a
ngo
r
,
M
a
l
a
y
s
i
a
.
E
m
a
il:
liz
a
j
a
m
i
l@
c
o
m
p
u
te
r
.
o
r
g
1.
I
NT
RO
D
UCT
I
O
N
C
o
n
vo
l
ut
i
o
na
l
N
e
ur
a
l
N
e
t
w
o
r
ks
(
C
N
N
)
ha
s
b
e
e
n
s
uc
c
e
s
s
f
ul
l
y
a
ppl
i
e
d i
n
m
a
ny
r
e
c
og
n
i
t
i
o
n
p
r
o
b
l
em
s
.
T
h
e co
n
cep
t
u
al
ar
c
h
i
t
ect
u
r
e o
f
C
N
N
i
s
i
n
s
p
i
r
ed
b
y
[1
]
’
s
s
e
m
in
a
l
w
o
r
k
o
n
t
h
e
c
a
t’
s
s
tr
ia
te
c
o
r
te
x
cal
l
ed
r
ecep
t
i
v
e f
i
el
d
.
L
at
er
,
F
u
k
u
s
h
i
m
a
[2
]
,
e
xp
l
a
i
ne
d
t
he
N
e
o
c
o
gni
t
r
o
n,
w
hi
c
h d
e
f
i
ne
s
t
he
l
a
y
e
r
w
i
s
e
s
t
r
u
ct
u
r
e o
f
n
eu
r
al
n
e
t
w
o
r
k
s
an
d
ex
p
l
ai
n
s
t
h
e
s
p
at
i
al
i
n
v
ar
i
an
ce c
h
ar
act
er
i
s
t
i
c
o
f
s
i
m
p
l
e
cel
l
s
an
d
co
mp
l
e
x
cel
l
s
o
f
v
i
s
u
al
p
r
i
m
ar
y
co
r
t
ex
.
L
e
C
u
n
i
n
t
r
o
d
u
ced
t
h
e
s
t
r
u
ct
u
r
e o
f
C
N
N
f
o
r
f
ace
an
d
d
i
g
i
t
r
eco
g
n
i
t
i
o
n
[3
-
5]
w
h
ic
h
d
e
m
o
n
s
tr
a
te
d
b
e
tte
r
r
e
c
o
g
n
itio
n
r
e
s
ul
t
s
t
h
a
n
pr
oba
bi
l
i
t
y
de
ns
i
t
y
f
u
n
c
t
i
on
m
e
t
h
odol
og
i
e
s
s
u
c
h
a
s
G
au
s
s
i
a
n
B
a
y
es
i
a
n
ap
p
r
o
ach
es
an
d
G
au
s
s
i
a
n
M
i
x
t
u
r
e
m
o
d
e
l
s
.
T
h
ey
d
ev
e
l
o
p
ed
a m
u
l
t
i
-
la
y
e
r
a
r
tif
ic
ia
l
n
e
u
r
a
l
n
et
w
o
r
k
cal
l
ed
L
eN
et
-
5
w
h
i
c
h
c
o
ul
d
c
l
a
s
s
i
f
y ha
nd
w
r
i
t
t
e
n
d
i
gi
t
s
.
L
i
ke
o
t
he
r
ne
ur
a
l
ne
t
w
o
r
k
s
,
L
eN
et
-
5
ha
s
m
u
l
t
i
p
l
e
l
a
y
er
s
an
d
can
b
e
t
r
ai
n
ed
w
i
t
h
t
h
e
b
ack
p
r
o
p
ag
at
i
o
n
al
g
o
r
i
t
h
m
[
6
]
.
I
t
can
o
b
t
ai
n
ef
f
ect
i
v
e
r
e
p
r
e
s
e
n
ta
tio
n
s
o
f
t
h
e
o
r
ig
in
a
l
i
m
a
g
e
,
w
h
ic
h
m
a
k
e
s
it p
o
s
s
i
b
le
to
r
e
c
o
g
n
iz
e
v
is
u
a
l p
a
tte
r
n
s
d
ir
e
c
tl
y
f
r
o
m
r
a
w
p
ix
e
ls
w
it
h
litt
le
-
to
-
n
on
e
pr
e
pr
o
ces
s
i
n
g
.
G
e
n
er
al
l
y
,
C
N
N
s
a
r
e co
n
s
t
r
u
c
t
ed
b
y
s
t
ack
i
n
g
i
n
t
er
w
ea
v
ed
l
a
y
er
s
o
f
t
w
o t
y
pe
s
:
c
onv
ol
u
t
i
on
a
l
l
a
y
e
r
s
a
n
d pool
i
ng
(
s
u
bs
a
m
pl
i
n
g
)
l
a
y
e
r
s
[7
]
.
T
h
e
pool
i
n
g
l
a
y
e
r
pl
a
y
s
a
n
i
m
por
t
a
nt
r
o
le
in
C
N
N
s
s
i
n
c
e
i
t i
s
m
a
in
l
y
r
e
s
p
o
n
s
ib
le
f
o
r
th
e
in
v
a
r
ia
n
c
e
to
d
a
ta
v
a
r
ia
tio
n
a
n
d
p
e
r
tu
r
b
a
tio
n
.
T
h
e
p
o
o
lin
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nd
o
ne
s
i
a
n J
E
l
e
c
E
ng
&
C
o
m
p
S
c
i
I
SSN
:
2502
-
4752
C
an C
onv
ol
ut
i
on N
e
ur
al
N
e
t
w
or
k
(
C
N
N
)
T
r
i
um
ph i
n E
ar
R
e
c
ogni
t
i
on
of
U
ni
f
or
m
…
(
Nu
r
s
u
ri
a
ti J
a
m
il
)
559
l
a
y
e
r
c
a
n
be
obt
a
i
n
e
d t
h
r
ou
g
h pool
i
n
g
ope
r
a
t
i
on
w
h
i
c
h
t
y
pi
c
a
l
l
y
c
on
t
a
i
n
s
t
w
o s
t
e
ps
.
F
i
r
s
t
l
y
,
pool
i
n
g
ope
r
a
t
or
s
can
s
t
h
e
f
eat
u
r
e
m
ap
an
d
ag
g
r
eg
at
es
i
n
f
o
r
m
at
i
o
n
(
r
es
p
o
n
s
e
s
)
w
i
t
h
i
n
each
l
o
cal
r
eg
i
o
n
[8
]
.
D
e
e
p
c
o
nvo
l
ut
i
o
n
ne
ur
a
l
ne
t
w
o
r
k
ha
s
s
ho
w
n
t
he
t
r
e
nd
o
f
d
o
m
i
na
t
i
ng
c
o
m
p
ut
e
r
vi
s
i
o
n
’
s
w
o
r
l
d
b
y
t
r
i
u
m
p
h
i
n
g
o
v
er
t
r
ad
i
t
i
o
n
al
m
ach
i
n
e
i
n
t
e
l
l
i
g
en
ce ap
p
r
o
ach
e
s
.
I
t
h
as
e
x
cel
l
ed
i
n
i
m
ag
e c
l
as
s
i
f
i
cat
i
o
n
[
9
,
10]
,
c
h
a
r
a
c
t
e
r
r
e
c
ogn
i
t
i
on
[
11]
a
nd
m
a
ny
m
or
e
.
T
h
e
r
obus
t
n
e
s
s
of
C
N
N
h
a
s
a
l
s
o
be
e
n
s
h
ow
n
i
n
w
or
k
b
y
[
12]
w
h
er
e t
h
e
l
a
y
er
s
o
f
C
N
N
ar
e
s
t
u
d
i
ed
t
o
b
et
t
er
u
n
d
er
s
t
an
d
h
o
w
v
i
e
w
p
o
i
n
t
i
n
v
ar
i
a
n
ce i
s
ac
h
i
ev
ed
.
T
h
e d
at
as
et
s
u
s
ed
ar
e R
G
B
-
D
d
at
as
et
[
1
3
]
an
d
P
as
cal
3
D
+ d
at
as
et
[
1
4
]
as
b
o
t
h
d
at
as
et
s
co
n
t
ai
n
m
u
l
t
i
v
i
e
w
t
ab
l
et
o
p
o
b
j
ect
s
a
n
d i
m
a
g
e
s
o
f
obj
e
c
t
c
a
t
e
g
or
i
e
s
e
xh
i
bi
t
i
n
g
h
i
gh
v
a
r
i
a
bi
l
i
t
y
,
c
a
pt
u
r
e
d i
n
un
c
ont
r
ol
l
e
d s
e
t
t
i
ng
s
a
n
d un
de
r
m
a
ny
di
f
f
e
r
e
n
t
pos
e
s
.
W
h
e
n
us
e
d on
pr
e
-
t
r
ai
n
ed
C
N
N
s
,
t
h
e n
et
w
o
r
k
cap
t
u
r
es
r
ep
r
es
en
t
at
i
o
n
s
t
ha
t
hi
g
hl
y
p
r
e
s
e
r
ve
t
he
m
a
n
i
f
o
l
d
s
t
r
uc
t
ur
e
a
t
m
o
s
t
o
f
t
he
ne
t
w
o
r
k l
a
ye
r
s
,
i
nc
l
ud
i
ng
t
he
f
u
ll
y
c
o
n
n
e
c
te
d
la
y
e
r
s
.
I
t is
a
ls
o
n
o
te
d
th
a
t
th
e
la
te
r
la
y
e
r
s
s
u
c
h
a
s
P
o
o
l 5
s
h
o
w
s
b
e
tte
r
r
e
p
r
e
s
e
n
ta
tio
n
f
o
r
th
e
v
ie
w
-
m
a
ni
f
o
l
d
t
ha
n e
a
r
l
y l
a
y
e
r
s
l
i
ke
P
o
o
l
1.
W
he
n
t
he
ne
t
w
o
r
k i
s
f
i
ne
-
t
un
e
d,
v
e
r
y
g
ood pos
e
e
s
t
i
m
a
t
i
on
pe
r
f
or
m
a
n
c
e
i
s
a
c
hi
e
v
e
d
.
I
n
[
1
5]
,
C
N
N
’
s
t
o
l
er
an
ce t
o
i
m
a
g
e
v
ar
i
at
i
o
n
s
s
u
c
h
a
s
t
r
an
s
l
at
i
o
n
,
s
cal
e,
p
o
s
e an
d
i
l
l
u
m
i
n
at
i
o
n
ar
e i
n
v
e
s
t
i
g
at
ed
u
s
i
n
g
a l
ar
g
e
-
s
cal
e s
y
n
t
h
et
i
c d
at
as
et
.
T
h
ei
r
d
at
as
et
s
co
n
s
i
s
t
s
o
f
i
m
a
g
e o
b
j
ect
s
o
f
1
6
cat
eg
o
r
i
es
,
8
r
o
t
at
i
o
n
an
g
l
es
,
1
1
ca
m
er
as
o
n
a s
e
m
i
ci
r
c
u
l
ar
ar
ch
f
o
r
r
an
d
o
m
v
i
e
w
p
o
i
n
t
s
an
d
d
i
s
t
an
ces
,
5
l
i
g
h
t
i
n
g
co
n
d
i
t
i
o
n
s
,
3
f
o
cu
s
l
ev
el
s
,
v
a
r
i
e
t
y
of
ba
c
kg
r
oun
ds
g
e
n
e
r
a
t
i
ng
ov
e
r
20
m
i
l
l
i
on
i
m
a
g
e
s
i
n
t
ot
a
l
.
A
pr
e
-
t
r
a
i
ne
d
ne
t
w
o
r
k,
na
m
e
l
y A
l
e
xN
e
t
C
N
N
i
s
u
s
ed
t
o
d
i
s
co
v
er
t
h
e
ex
p
r
es
s
i
v
e cap
ab
i
l
i
t
i
e
s
o
f
t
h
e p
o
o
l
5
an
d
f
c7
l
ay
er
s
f
o
r
o
b
j
ect
an
d
p
ar
am
et
er
pr
e
di
c
t
i
on
s
.
R
e
s
u
l
t
s
s
h
o
w
e
d t
ha
t
bot
h
f
c
7 a
n
d pool
5 pr
odu
c
e
d a
c
c
u
r
a
t
e
r
e
s
u
l
t
s
f
or
c
l
a
s
s
i
f
i
c
a
t
i
on
i
m
pl
y
i
ng
t
h
a
t
t
he
s
e
l
a
ye
r
s
ha
ve
go
o
d
d
i
s
c
r
i
m
i
n
at
i
v
e p
o
w
er
f
o
r
o
b
j
ect
r
eco
g
n
i
t
i
o
n
.
A
s
f
o
r
p
ar
am
e
t
er
p
r
e
d
i
ct
i
o
n
s
,
t
h
e
accu
r
aci
es
f
o
r
l
i
g
h
t
i
n
g
,
r
o
t
at
i
o
n
,
an
d
ca
m
er
a ar
e 1
0
0
%
,
7
7
%
,
an
d
6
2
%
.
,
r
es
p
ect
i
v
el
y
.
T
h
i
s
i
m
p
l
i
es
t
h
at
t
h
e
la
y
e
r
s
h
a
v
e
d
i
f
f
ic
u
ltie
s
m
o
s
t
i
n
p
r
e
d
ic
tin
g
c
a
m
e
r
a
v
ie
w
,
w
h
ile
lig
h
ti
n
g
v
a
r
ia
tio
n
s
p
o
s
e
v
e
r
y
little
p
r
e
d
ic
tio
n
d
i
f
f
i
c
u
l
t
y
.
A
s
c
h
an
g
i
n
g
ca
m
er
a v
i
e
w
s
al
t
er
p
r
o
d
u
ce g
eo
m
et
r
i
c v
ar
i
a
n
t
s
,
p
r
ed
i
ct
i
o
n
t
as
k
b
eco
m
e
m
o
r
e
c
o
m
p
l
i
c
a
t
e
d
.
T
hi
s
f
i
nd
i
ng
i
s
f
ur
t
he
r
p
r
o
ve
n
w
h
e
n a
n a
s
s
e
s
s
m
e
n
t
w
a
s
d
o
ne
o
n t
he
ne
t
w
o
r
k t
o
l
e
a
r
n
t
he
p
o
w
e
r
o
f C
N
N
i
n
t
r
a
n
s
fe
r
r
i
n
g
t
h
e
l
ear
n
ed
p
ar
am
e
t
er
o
v
er
o
n
e o
b
j
ect
cat
eg
o
r
y
t
o
a
n
o
t
h
er
.
L
i
g
h
t
i
n
g
p
ar
a
m
e
t
er
i
s
eas
i
l
y
t
r
an
s
f
er
r
ed
f
r
o
m
s
ee
n
o
b
j
ect
s
t
o
u
n
s
ee
n
o
b
j
ect
,
w
h
i
l
e
v
ar
i
at
i
o
n
s
i
n
ca
m
er
a
v
i
e
w
s
d
eg
r
ad
ed
t
h
e
pe
r
f
or
m
a
n
c
e
of
t
h
e
n
e
t
w
or
k
f
or
u
n
s
e
e
n
obj
e
c
t
s
.
A
s
l
i
gh
t
i
ng v
a
r
i
a
t
i
on
s
d
o
n
o
t
t
r
an
s
f
o
r
m
t
h
e g
eo
m
et
r
i
c s
h
ap
es
o
f
t
h
e s
ee
n
o
b
j
ect
,
i
t
h
as
t
h
e s
i
m
p
l
es
t
k
n
o
w
l
ed
g
e t
o
b
e t
r
an
s
f
er
r
ed
o
n
u
n
s
ee
n
cat
eg
o
r
i
es
.
E
v
en
t
h
o
u
g
h
v
i
e
w
p
o
i
n
t
o
r
cam
er
a v
i
e
w
i
n
v
ar
i
an
ce
r
e
m
ai
n
s
a ch
al
l
en
g
e i
n
a
n
y
o
b
j
ect
r
e
c
og
n
i
t
i
o
n
[
17]
,
w
e
onl
y
c
o
n
s
i
de
r
i
l
l
um
i
n
a
tio
n
in
v
a
r
ia
n
t
in
t
h
is
p
a
p
e
r
.
D
e
s
p
ite
th
e
f
a
c
t th
a
t
C
N
N
s
h
a
v
e
s
ho
w
n
t
o
ha
nd
l
e
i
l
l
u
m
i
na
t
i
o
n
i
nva
r
i
a
nt
f
o
r
o
b
j
e
c
t
r
e
c
o
gni
t
i
o
n,
w
e
b
e
l
i
e
ve
t
ha
t
f
ur
t
he
r
s
t
ud
y
ne
e
d
t
o
b
e
d
o
ne
f
or
i
m
a
g
e
s
o
f
un
i
f
or
m
i
l
l
um
i
na
t
i
on
i
nv
a
r
i
a
n
t
a
s
m
os
t
w
or
k
f
oc
u
s
on
l
y
on
n
on
-
u
ni
f
or
m
i
n
v
a
ri
a
n
t
[1
8
].
F
i
g
u
re
1
a
n
d 2 i
l
l
us
t
r
a
t
e
pr
e
v
i
ous
w
or
k [
15,
1
6]
on
i
l
l
um
i
n
a
t
i
on
i
nv
a
r
i
a
n
t
w
he
r
e
t
h
e
i
m
a
g
e
da
t
a
s
e
t
s
u
s
e
d a
r
e
on
l
y
n
o
n
-
u
n
i
f
o
r
m
i
llu
m
i
n
a
tio
n
in
v
a
r
ia
n
t
c
a
u
s
e
d
b
y
c
h
a
n
g
e
s
i
n
lig
h
ti
n
g
in
te
n
s
it
y
,
d
ir
e
c
tio
n
a
n
d
s
p
e
c
tr
u
m
.
F
i
gu
r
e
1.
A
boa
t
u
n
de
r
5
d
if
f
e
r
e
n
t ill
u
m
in
a
tio
n
s
[
1
5
]
O
n t
he
o
t
he
r
ha
nd
,
V
o
ni
ka
ki
s
e
t
a
l
.
[
1
8
]
c
o
ns
t
r
uc
t
e
d
a
n
e
w
b
en
ch
m
ar
k
d
at
as
et
n
a
m
ed
P
h
o
s
f
eat
u
r
i
n
g
a g
r
eat
er
v
ar
i
et
y
o
f
i
m
a
g
i
n
g
co
n
d
i
t
i
o
n
s
,
co
m
p
ar
ed
t
o
ex
i
s
t
i
n
g
d
at
ab
as
e
s
;
co
n
t
a
i
n
i
n
g
i
m
ag
es
cap
t
u
r
ed
b
o
t
h
und
e
r
u
ni
f
o
r
m
a
nd
no
n
-
u
n
i
f
or
m
i
l
l
um
i
n
a
t
i
on
.
P
h
os
c
o
m
pr
i
s
e
s
15 s
c
e
n
e
s
c
a
pt
u
r
e
d un
de
r
d
if
f
e
r
e
n
t ill
u
m
i
n
a
tio
n
c
on
di
t
i
on
s
.
E
a
c
h s
c
e
n
e
c
ont
a
i
n
s
9 i
m
a
g
e
s
t
a
k
e
n
un
de
r
di
s
s
i
m
i
l
a
r
un
i
f
or
m
i
l
l
u
m
i
na
t
i
o
ns
,
a
nd
6
i
m
a
ge
s
u
nd
e
r
di
f
f
e
r
e
n
t
de
g
r
e
e
s
of
n
on
-
un
i
f
or
m
i
l
l
um
i
n
a
t
i
o
n
s
.
F
i
gu
r
e
3
de
m
on
s
t
r
at
es
a
n
e
x
a
m
p
l
e o
f
a
s
cen
e
i
n
P
h
o
s
d
a
ta
b
a
s
e
.
U
n
if
o
r
m
il
lu
m
i
n
a
tio
n
is
a
tta
i
n
e
d
b
y
u
s
i
n
g
s
e
v
e
r
a
l d
if
f
i
u
s
iv
e
li
g
h
t s
o
u
r
c
e
s
e
v
e
n
l
y
d
is
tr
ib
u
te
d
a
r
o
u
n
d
t
h
e
obj
e
c
t
s
.
U
s
i
ng
t
h
e
s
hu
t
t
e
r
s
pe
e
d,
f
ou
r
ov
e
r
e
x
pos
e
d (
+
)
a
n
d f
ou
r
un
de
r
e
x
pos
e
d (
-
)
i
m
ag
es
ar
e
g
en
er
at
ed
fr
o
m
th
e
c
o
r
r
e
c
t e
x
p
o
s
u
r
e
i
m
a
g
e
.
A
s
tr
o
n
g
d
ir
e
c
tio
n
a
l li
g
h
t s
o
u
r
c
e
w
i
th
d
i
f
f
e
r
e
n
t s
tr
e
n
g
t
h
s
is
p
r
o
j
e
c
te
d
to
th
e
ob
j
e
c
t
s
pr
odu
c
i
n
g
6 n
on
-
uni
f
o
r
m
i
l
l
u
m
i
na
t
i
o
n i
m
a
ge
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
25
02
-
4752
I
nd
o
ne
s
i
a
n J
E
l
e
c
E
ng
&
C
o
m
p
S
c
i
,
V
o
l.
11
, N
o
.
2
,
A
ug
us
t
2018
:
5
58
–
566
56
0
F
i
g
ur
e
2
.
R
G
B
i
m
a
ge
s
t
a
ke
n a
t
t
he
s
a
m
e
l
o
c
a
t
i
o
n e
ve
r
y
ho
ur
o
f
a
2
4
-
h
ou
r
pe
r
i
od
[
16
]
F
i
g
ur
e
3
.
R
G
B
i
m
a
ge
s
t
a
ke
n a
t
t
he
s
a
m
e
l
o
ca
t
i
o
n e
ve
r
y
ho
ur
o
f
a
2
4
-
h
ou
r
pe
r
i
od
[
16
]
.
E
v
en
t
h
o
u
g
h
,
P
h
o
s
i
m
a
g
e d
at
as
et
s
ar
e
v
al
u
ab
l
e an
d
i
s
f
r
eel
y
acces
s
i
b
l
e
f
o
r
p
u
b
l
i
c
u
s
e,
t
h
e u
n
i
f
o
r
m
i
l
l
um
i
n
a
t
i
on i
s
i
n
a
c
o
n
t
r
ol
l
e
d
e
n
vi
r
onm
e
n
t
.
U
n
de
r
a
n
d
ove
r
-
e
x
pos
e
d
pi
x
e
l
s
i
n
R
G
B
i
m
ag
es
ar
e
t
h
e
m
aj
o
r
s
o
u
r
c
e
o
f
a
r
te
f
a
c
ts
i
n
th
e
il
lu
m
i
n
a
tio
n
i
n
v
a
r
ia
n
t,
p
a
r
tic
u
la
r
l
y
d
u
r
in
g
d
a
y
li
g
h
t h
o
u
r
s
[
1
6
]
.
T
h
e
r
e
f
o
r
e
,
o
u
r
f
ir
s
t
c
o
n
tr
ib
u
tio
n
i
n
th
is
p
a
p
e
r
is
th
e
c
o
lle
c
tio
n
o
f
u
n
if
o
r
m
i
llu
m
i
n
a
tio
n
i
m
a
g
e
s
u
n
d
e
r
n
a
t
u
r
a
l s
o
u
r
c
e
illu
m
i
n
a
tio
n
s
.
O
ur
uni
f
o
r
m
i
l
l
u
m
i
na
t
i
o
n
i
m
a
ge
s
a
r
e
e
a
r
i
m
a
ge
s
c
a
p
t
ur
e
d
i
n
unc
o
nt
r
o
l
l
e
d
i
nd
o
o
r
a
nd
o
ut
d
o
o
r
e
nvi
r
o
n
m
e
nt
.
W
e ch
o
o
s
e ear
b
i
o
m
e
t
r
i
cs
d
o
m
ai
n
as
ear
r
eco
g
n
i
t
i
o
n
u
s
i
n
g
d
eep
C
N
N
i
s
s
t
i
l
l
s
car
ce d
u
e t
o
l
ack
o
f
l
ar
g
e
-
s
cal
e
d
a
ta
s
e
ts
[
1
9
]
.
A
t t
h
e
ti
m
e
o
f
w
r
iti
n
g
,
o
n
l
y
f
o
u
r
pu
bl
i
s
h
e
d
w
or
k [
19]
-
[
22]
of
e
m
pl
o
y
i
ng
de
e
p l
e
a
r
n
i
ng
f
or
e
a
r
r
eco
g
n
i
t
i
o
n
ar
e
f
o
u
n
d
,
w
i
t
h
t
w
o
p
ap
er
s
f
r
o
m
t
h
e
s
a
m
e t
ea
m
.
A
l
t
h
o
u
g
h
t
h
e
y
al
s
o
u
t
i
l
i
ze
d
C
N
N
ar
c
h
i
t
ect
u
r
e,
b
u
t
t
h
e d
at
as
et
s
u
s
ed
ar
e u
n
d
er
co
n
t
r
o
l
l
ed
en
v
i
r
o
n
m
e
n
t
o
r
ea
r
i
m
a
g
es
w
i
t
h
n
o
n
-
un
i
f
or
m
ill
u
m
in
a
t
io
n
c
a
p
tu
r
e
d
i
n
t
h
e
w
i
l
d
.
C
o
n
s
eq
u
e
n
t
l
y
,
t
h
i
s
p
ap
er
i
n
t
en
d
s
t
o
i
n
v
e
s
t
i
g
at
e t
h
e accu
r
ac
y
o
f
C
N
N
f
o
r
ear
b
i
o
m
et
r
i
c
s
cap
t
u
r
ed
und
e
r
u
ni
f
o
r
m
i
l
l
u
m
i
na
t
i
o
n i
n
va
r
i
a
nt
e
n
vi
r
o
n
m
e
nt
.
2.
T
H
E
UN
I
F
O
RM
I
L
L
UM
I
NANT
I
N
VARI
ANT
E
AR I
M
AG
E
DAT
AS
E
T
I
llu
m
in
a
tio
n
is
t
h
e
a
m
o
un
t
o
f
l
i
g
ht
p
r
e
s
e
nt
i
n a
n a
r
e
a
[
2
3
]
a
nd
i
s
m
e
a
s
ur
e
d
i
n l
ux (
l
x)
o
r
l
u
m
e
n (
l
m
)
p
er
s
q
u
ar
e
m
et
er
.
L
u
x
,
al
s
o
k
n
o
w
n
as
l
u
m
i
n
an
ce i
s
a
m
eas
u
r
e
m
en
t
o
f
b
r
i
g
h
t
n
es
s
m
eas
u
r
ed
u
s
i
n
g
a l
u
x
m
et
er
.
T
h
e u
n
i
q
u
e
n
es
s
o
f
o
u
r
d
at
as
et
s
i
s
t
h
at
t
h
e ear
i
m
a
g
es
ar
e acq
u
i
r
ed
i
n
t
h
e
i
n
door
a
n
d ou
t
door
e
n
v
i
r
onm
e
n
t
s
,
a
n
d
illu
m
i
n
a
tio
n
s
a
r
e
o
b
j
e
c
tiv
e
l
y
q
u
a
n
ti
f
ie
d
u
s
i
n
g
l
u
x
m
e
te
r
a
t a
c
tu
a
l lo
c
a
tio
n
s
.
T
h
e
a
c
q
u
is
itio
n
p
r
o
c
e
s
s
s
t
ar
t
ed
w
i
t
h
t
h
e
s
u
b
j
ect
s
eat
e
d
o
n
a p
r
o
v
i
d
ed
ch
ai
r
.
T
h
e s
u
b
j
ect
'
s
ear
i
s
f
aced
t
o
w
ar
d
s
t
h
e ca
m
er
a,
an
d
t
h
e
ca
m
er
a i
s
s
et
t
o
b
e at
t
h
e
s
a
m
e l
ev
el
w
i
t
h
t
h
e ear
s
.
T
h
e
n
,
t
h
e i
l
l
u
m
i
n
at
i
o
n
i
s
m
eas
u
r
ed
u
s
i
n
g
l
u
x
m
et
er
.
T
h
e
l
u
x
m
et
er
s
e
n
s
o
r
i
s
p
l
aced
d
i
r
ect
l
y
u
n
d
er
t
h
e ear
t
o
en
s
u
r
e an
acc
u
r
at
e r
ead
i
n
g
.
D
et
ai
l
s
o
f
t
h
e i
m
ag
e
a
c
qu
i
s
i
t
i
o
n
c
a
n
be
f
oun
d i
n
o
u
r
pr
e
v
i
ous
w
or
k [
24]
.
F
or
t
he
un
i
f
or
m
i
l
l
um
i
n
a
t
i
o
n
v
a
r
i
a
nc
e
,
w
e
m
a
n
a
g
e
d t
o
c
o
lle
c
t e
a
r
i
m
a
g
e
s
w
it
h
ill
u
m
i
n
a
tio
n
s
r
a
n
g
i
n
g
f
r
o
m
2
lu
x
to
1
0
7
0
0
lu
x
f
r
o
m
5
5
s
u
b
j
ect
s
.
T
en
ear
i
m
a
g
es
ar
e
c
o
l
l
e
c
t
e
d
f
r
o
m
e
a
c
h
s
ub
j
e
c
t
,
w
i
t
h
va
r
i
o
us
l
u
x
va
l
ue
s
d
e
p
e
nd
i
ng
o
n
t
he
s
ur
r
o
und
i
n
g
l
i
g
ht
i
ng.
T
he
f
i
na
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nd
o
ne
s
i
a
n J
E
l
e
c
E
ng
&
C
o
m
p
S
c
i
I
SSN
:
2502
-
4752
C
an C
onv
ol
ut
i
on N
e
ur
al
N
e
t
w
or
k
(
C
N
N
)
T
r
i
um
ph i
n E
ar
R
e
c
ogni
t
i
on
of
U
ni
f
or
m
…
(
Nu
r
s
u
ri
a
ti J
a
m
il
)
561
c
om
pi
l
a
t
i
on
o
f
e
a
r
da
t
a
s
e
t
c
o
n
s
i
s
t
s
o
f
a
t
ot
a
l
o
f
1,
100 i
m
a
g
e
s
c
o
m
pr
i
s
i
ng
550 l
e
f
t
a
n
d
550 r
i
g
h
t
i
m
a
g
e
s
.
T
ab
l
e
1
l
i
s
t
s
t
h
e co
l
l
ect
i
o
n
o
f
ear
i
m
a
g
e
s
i
n
d
i
f
f
er
en
t
l
u
x
f
o
r
b
o
t
h
l
ef
t
a
n
d
r
i
g
h
t
ear
i
m
ag
e
s
.
M
ean
w
h
i
l
e,
F
i
gu
r
e
4 s
h
o
w
s
s
o
m
e
e
x
a
m
pl
e
s
of
t
h
e ear
i
m
ag
e
s
i
n
s
el
ec
t
ed
l
u
x
v
al
u
e
s
.
T
h
e r
es
o
l
u
t
i
o
n
f
o
r
t
h
e ear
i
m
ag
e
s
ar
e
2048 x
1
056
pi
x
e
l
s
.
T
ab
l
e 1
.
L
u
x
v
ar
i
at
i
o
n
s
o
f
t
h
e ear
i
m
ag
e d
at
as
et
Lu
x
Le
f
t
R
i
gh
t
To
t
a
l
Lu
x
Le
f
t
R
i
gh
t
To
t
a
l
Lu
x
Le
f
t
R
i
gh
t
To
t
a
l
1
-
1
0
1
3
0
1
0
6
2
3
6
1
0
1
-
2
00
5
2
5
0
1
0
2
2
0
0
1
-
3
0
0
0
1
6
1
4
3
0
1
1
-
2
0
5
4
8
0
1
3
4
2
0
1
-
3
00
3
4
3
2
6
6
3
0
0
1
-
4
0
0
0
4
6
1
0
2
1
-
3
0
3
2
2
8
6
0
3
0
1
-
4
00
4
8
5
0
9
8
4
0
0
1
-
5
0
0
0
8
8
1
6
3
1
-
4
0
6
6
1
2
4
0
1
-
5
00
2
8
1
0
5
0
0
1
-
6
0
0
0
6
6
1
2
4
1
-
5
0
3
0
3
6
6
6
5
0
1
-
6
00
4
6
1
0
6
0
0
1
-
7
0
0
0
0
2
2
5
1
-
6
0
1
4
8
2
2
6
0
1
-
7
00
6
4
1
0
7
0
0
1
-
8
0
0
0
6
2
8
6
1
-
7
0
8
4
1
2
7
0
1
-
8
00
4
4
8
8
0
0
1
-
9
0
0
0
2
2
4
7
1
-
8
0
1
0
1
8
2
8
8
0
1
-
9
00
6
4
1
0
9
0
0
1
-
1
0
0
00
0
0
0
8
1
-
9
0
2
2
1
6
3
8
9
0
1
-
1
00
0
2
6
8
1
0
0
0
1
-
1
1
0
0
0
0
2
2
9
1
-
1
0
0
1
2
1
4
2
6
1
0
0
1
-
2
0
0
0
3
2
2
8
6
0
T
o
t
a
l
5
5
0
5
5
0
1
1
0
0
5
l
u
x
1
8
l
u
x
3
3
l
u
x
5
0
l
u
x
8
5
l
u
x
1
0
0
l
u
x
3
0
0
l
u
x
5
5
0
l
u
x
9
6
0
l
u
x
1
1
8
0
l
u
x
2
1
8
0
l
u
x
5
1
1
0
l
u
x
F
i
g
ur
e
4
.
O
u
r
u
n
if
o
r
m
ill
u
m
in
a
tio
n
-
i
n
v
ar
i
an
t
ear
i
m
a
g
e d
at
as
et
D
a
ta
a
u
g
m
e
n
ta
tio
n
u
s
i
n
g
r
o
ta
tio
n
a
t e
v
e
r
y
5
o
i
s
ap
p
l
i
ed
t
o
t
h
e ear
i
m
ag
e
s
t
o
g
en
er
at
e ad
eq
u
at
e d
at
a
t
r
a
i
ni
ng a
nd
t
e
s
t
i
n
g
f
o
r
t
h
e C
N
N
m
o
d
el
.
A
f
t
er
d
at
a au
g
m
en
t
at
i
o
n
,
a t
o
t
al
o
f
2
5
,
3
0
0
i
m
a
g
e
s
ar
e g
en
er
at
ed
an
d
f
o
r
each
s
u
b
j
ect
,
w
e c
h
o
s
e ap
p
r
o
x
i
m
at
el
y
8
8
.
3
%
(
i
.
e.
4
1
0
i
m
ag
e
s
p
er
s
u
b
j
ect
)
t
o
b
e
u
s
ed
as
t
r
ai
n
i
n
g
d
at
as
et
a
n
d t
h
e
r
e
m
a
i
n
i
ng
11.
7%
(
50 i
m
a
g
e
s
pe
r
s
u
bj
e
c
t
)
i
s
u
s
e
d f
or
v
a
lid
a
tio
n
a
n
d
te
s
ti
n
g
d
a
ta
s
e
ts
.
3.
P
RO
P
O
S
E
D CNN M
O
DE
L
I
n
t
hi
s
pa
pe
r
,
w
e
pr
e
s
e
n
t
t
h
e
u
s
e
o
f
de
e
p c
on
v
ol
u
t
i
on
a
l
n
e
ur
a
l
n
e
t
w
or
ks
(
C
N
N
)
t
o i
de
n
t
i
f
y
a
pe
r
s
on
b
as
ed
o
n
ear
b
i
o
m
et
r
i
cs
.
O
n
t
h
e b
as
i
s
o
f
an
an
al
y
s
i
s
o
f
s
t
r
u
ct
u
r
e an
d
p
ar
am
et
er
s
i
n
C
N
N
,
t
h
e g
r
ad
i
en
t
-
d
es
ce
n
t
a
l
g
or
i
t
hm
c
a
n
be
a
ppl
i
e
d t
o t
r
a
i
n
C
N
N
.
A
t
ot
a
l
of
25000 e
a
r
i
m
a
g
e
s
a
r
e
r
e
s
i
z
e
d t
o 125x
125 t
o c
or
r
e
l
a
t
e
w
i
t
h
t
h
e s
i
ze o
f
t
h
e i
n
p
u
t
l
a
y
er
,
an
d
t
h
en
t
h
e r
es
i
zed
i
m
ag
e
s
ar
e u
s
ed
t
o
t
r
ai
n
t
h
e
C
N
N
m
o
d
el
.
T
h
i
s
C
N
N
s
m
o
d
el
can
i
m
p
r
o
v
e
t
h
e
co
n
v
er
g
en
ce
s
p
ee
d
w
hi
l
e
t
r
a
i
ni
n
g
t
he
p
a
r
a
m
e
t
e
r
s
i
n
C
N
N
s
,
a
nd
o
b
t
a
i
n
hi
g
he
r
r
e
c
o
gni
t
i
o
n
accu
r
ac
y
co
m
p
ar
ed
t
o
co
n
v
en
t
i
o
n
al
n
e
u
r
al
n
et
w
o
r
k
o
r
s
u
p
p
o
r
t
v
ect
o
m
ach
i
n
e.
T
h
e
ad
v
an
t
a
g
es
o
f
t
h
e
p
r
o
p
o
s
ed
C
N
N
ar
e t
h
at
i
m
a
g
e
s
can
b
e i
n
p
u
t
d
i
r
ect
l
y
t
o
t
h
e
m
o
d
el
.
F
i
g
u
r
e 5
d
em
o
n
s
t
r
at
es
t
h
e
pr
opos
e
d
C
N
N
m
o
d
el
a
n
d
t
h
e l
a
y
er
s
ar
e s
u
b
s
e
q
u
en
t
l
y
d
es
cr
i
b
ed
i
n
t
h
e
f
o
l
l
o
w
i
n
g
s
u
b
s
ect
i
o
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
25
02
-
4752
I
nd
o
ne
s
i
a
n J
E
l
e
c
E
ng
&
C
o
m
p
S
c
i
,
V
o
l.
11
, N
o
.
2
,
A
ug
us
t
2018
:
5
58
–
566
562
3.
1.
C
on
v
ol
u
t
i
on
al
L
aye
r
s
F
i
r
st
,
an
i
m
ag
e i
n
p
u
t
l
a
y
er
i
n
p
u
t
s
i
m
a
g
es
t
o
a n
et
w
o
r
k
an
d
a
p
p
l
i
es
d
at
a n
o
r
m
al
i
zat
i
o
n
.
T
h
en
,
w
e ca
n
cl
as
s
i
f
y
i
m
a
g
es
f
r
o
m
a r
ed
u
ced
d
at
a
s
e
t
a
ppl
y
i
n
g
c
onv
ol
u
t
i
on
a
n
d pool
i
n
g
[
25]
.
T
h
e
d
e
ta
ils
o
f
th
e
f
ir
s
t
co
n
v
o
l
u
t
i
o
n
al
l
a
y
er
ar
e p
r
es
en
t
ed
i
n
T
ab
l
e 2
.
F
i
g
ur
e
5
.
O
u
r
pr
opos
e
d C
N
N
m
ode
l
T
ab
l
e
2
.
F
i
r
s
t
c
on
v
ol
u
t
i
o
n
a
l
l
a
y
e
r
of
ou
r
pr
opos
e
d C
N
N
La
y
e
r
T
y
p
e
L
a
y
er
D
et
a
i
l
I
n
p
u
t
i
m
a
g
e
(
1
2
5
,
1
2
5
,
1
)
F
il
t
e
r
s
1
6
F
il
t
e
r
s
i
z
e
3
*
3
N
u
m
b
e
r o
f w
e
i
g
h
t
s
p
e
r fi
l
t
e
r
3
*
3
*
1=
9
N
u
m
b
er
o
f
p
a
r
a
m
et
er
s
i
n
t
h
e l
a
y
er
(
9
+
1
)
*
1
6
=
16
0
S
tr
id
e
1
N
u
m
b
e
r
o
f
n
e
u
r
o
n
s
i
n
e
a
c
h
f
e
a
t
u
r
e
m
a
p
(
1
25
-
3
+
2
*
1
)
/
1
+
1
=
1
2
5
T
o
t
a
l
n
u
m
b
e
r
o
f
n
e
u
r
o
n
s
i
n
t
h
e
l
a
y
e
r
1
2
5
*
12
5
*
1
6
=
2
50
0
0
0
S
eco
n
d
l
y
,
a b
at
ch
n
o
r
m
al
i
za
t
i
o
n
l
a
y
er
i
s
ad
d
ed
af
t
er
t
h
e co
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
t
o s
pe
e
d u
p
t
r
a
i
n
i
n
g
of
c
o
n
v
o
lu
tio
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
a
n
d
r
e
d
u
c
e
th
e
s
e
n
s
iti
v
it
y
to
n
e
t
w
o
r
k
in
it
ia
liz
a
tio
n
.
B
a
t
c
h
N
o
r
m
a
liz
a
tio
n
al
l
o
w
s
u
s
t
o
u
s
e
m
u
c
h
h
i
g
h
er
l
ear
n
i
n
g
r
at
e
s
a
n
d
b
e l
es
s
car
ef
u
l
ab
o
u
t
i
n
i
t
i
al
i
zat
i
o
n
.
I
t
al
s
o
act
s
a
s
a
r
eg
u
l
ar
i
zer
,
i
n
s
o
m
e cas
e
s
el
i
m
i
n
at
i
n
g
t
h
e n
eed
f
o
r
D
r
o
p
o
u
t
[
26]
.
T
h
ir
d
ly
,
R
e
c
ti
f
ie
d
L
i
n
e
a
r
U
n
its
(
R
e
L
U
)
is
u
s
ed
as
t
h
e act
i
v
at
i
o
n
f
u
n
ct
i
o
n
.
T
h
e R
e
L
U
l
a
y
er
p
er
f
o
r
m
s
a t
h
r
es
h
o
l
d
o
p
er
at
i
o
n
t
o
each
p
i
x
el
el
e
m
e
n
t
o
f
t
h
e
i
np
ut
a
s
s
e
e
n
i
n E
q
u
a
tio
n
1
[2
7
],
w
h
e
re
f
(x
)
i
s
t
he
i
m
a
ge
a
n
d
x
i
s
t
he
p
i
xe
l
e
l
e
m
e
nt
.
I
f
t
he
p
i
xe
l e
le
m
e
n
t,
x
is
le
s
s
t
h
a
n
z
e
r
o
it
w
ill b
e
r
e
s
e
t t
o
0
.
O
th
e
r
w
is
e
,
t
h
e
p
ix
e
l e
le
m
e
n
t r
e
m
a
i
n
s
a
s
is
.
(
)
=
,
≥
0
0
,
<
0
(
1
)
F
o
ur
t
hl
y,
a
m
a
x
pool
i
ng
l
a
y
e
r
pe
r
f
or
m
s
do
w
n
-
s
a
m
p
lin
g
b
y
d
iv
id
i
n
g
th
e
in
p
u
t in
to
r
e
c
ta
n
g
u
la
r
p
o
o
lin
g
r
e
gi
o
ns
,
a
nd
c
o
m
p
ut
i
n
g t
he
m
a
xi
m
u
m
o
f
e
a
c
h r
e
gi
o
n.
W
e
c
o
nf
i
g
ur
e
d
t
hi
s
l
a
ye
r
t
o
us
e
4
×
4
pool
i
n
g
r
e
g
i
o
n
s
i
z
e
a
n
d 2 f
or
t
h
e
s
t
r
i
de
.
T
h
i
s
s
t
r
u
c
t
u
r
e
i
s
r
e
pe
a
t
e
d 3 t
i
m
e
s
t
o f
or
m
3
f
ol
ds
o
f
l
a
y
e
r
s
.
T
h
e
m
a
x pool
i
n
g l
a
y
e
r
i
s
p
l
aced
b
et
w
ee
n
each
o
f
t
h
e t
w
o
f
o
l
d
s
as
i
l
l
u
s
t
r
at
ed
i
n
F
i
g
u
r
e
5.
T
a
b
le
3
s
h
o
w
s
t
h
e
la
y
e
r
d
e
ta
ils
o
f
t
h
e
s
e
c
o
n
d
c
o
n
v
o
l
u
tio
n
a
l
la
y
e
r
.
I
t
c
o
n
s
i
s
ts
o
f
3
2
f
ilt
e
r
s
o
f
s
iz
e
6
×
6
a
n
d
pa
ddi
n
g
o
f
1.
T
h
e
c
on
vol
u
t
i
on
a
l
l
a
y
e
r
i
s
f
u
r
t
h
e
r
f
ol
l
o
w
e
d
b
y
a
pool
i
ng
l
a
y
e
r
c
om
pr
i
s
i
n
g
2
×
2
p
o
o
lin
g
r
e
g
io
n
s
iz
e
a
n
d
2
f
o
r
th
e
s
tr
id
e
.
T
h
e
th
ir
d c
onv
ol
ut
i
on
a
l
l
a
y
e
r
c
ont
a
i
ns
64 f
i
l
t
e
r
s
a
n
d t
h
e
s
i
z
e
of
e
a
c
h
f
ilte
r
is
1
6
×
1
6
as
can
b
e s
ee
n
i
n
T
ab
l
e 4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nd
o
ne
s
i
a
n J
E
l
e
c
E
ng
&
C
o
m
p
S
c
i
I
SSN
:
2502
-
4752
C
an C
onv
ol
ut
i
on N
e
ur
al
N
e
t
w
or
k
(
C
N
N
)
T
r
i
um
ph i
n E
ar
R
e
c
ogni
t
i
on
of
U
ni
f
or
m
…
(
Nu
r
s
u
ri
a
ti J
a
m
il
)
563
T
ab
l
e
3
.
S
e
c
on
d c
on
v
ol
u
t
i
on
a
l
l
a
y
e
r
of
ou
r
pr
opos
e
d C
N
N
La
y
e
r
T
y
p
e
L
a
y
er
D
et
a
i
l
s
F
il
t
e
r
s
3
2
F
il
t
e
r
s
i
z
e
6
×
6
N
u
m
b
er
o
f
w
ei
g
h
t
s
p
er
f
i
l
t
er
6
×
6
×
1
=3
6
N
u
m
b
er
o
f
p
a
r
a
m
et
er
s
i
n
t
h
e l
a
y
er
(
3
6+
1
)
*
3
2
=
1
1
8
4
S
tr
id
e
1
N
u
m
b
e
r
o
f
n
e
u
r
o
n
s
i
n
e
a
c
h
f
e
a
t
u
r
e
m
a
p
(
1
25
-
3
2
+
2
*1
)
/
1
+
1
=
9
6
T
o
t
a
l
n
u
m
b
e
r
o
f
n
e
u
r
o
n
s
i
n
t
h
e
l
a
y
e
r
9
6
*
9
6
*
3
2
=
2
9
4
9
1
2
T
ab
l
e
4
.
T
h
i
r
d c
on
v
ol
u
t
i
on
a
l
l
a
y
e
r
of
ou
r
pr
opos
e
d C
N
N
La
y
e
r
T
y
p
e
L
a
y
er
D
et
a
i
l
s
F
il
t
e
r
s
6
4
F
il
t
e
r
s
i
z
e
1
6
×
1
6
N
u
m
b
er
o
f
w
ei
g
h
t
s
p
er
f
i
l
t
er
1
6
×
1
6
×
1
=
25
6
N
u
m
b
er
o
f
p
a
r
a
m
et
er
s
i
n
t
h
e l
a
y
er
(
2
5
6
+
1
)
*
6
4
=
1
6
4
4
8
S
tr
id
e
1
N
u
m
b
e
r
o
f
n
e
u
r
o
n
s
i
n
e
a
c
h
f
e
a
t
u
r
e
m
a
p
(
1
25
-
6
4
+
2
*1
)
/
1
+
1
=
6
4
T
o
t
a
l
n
u
m
b
e
r
o
f
n
e
u
r
o
n
s
i
n
t
h
e
l
a
y
e
r
6
4
*
6
4
*
6
4
=
2
6
2
1
4
4
3.
2
. F
u
lly
-
C
o
n
n
ect
ed
L
a
y
er
I
n
C
N
N
,
a
ll t
h
e
c
o
n
v
o
l
u
tio
n
a
l la
y
e
r
s
a
n
d
p
o
o
lin
g
la
y
e
r
s
a
r
e
f
o
llo
w
e
d
b
y
s
e
v
e
r
a
l f
u
ll
y
-
co
n
n
ect
ed
l
a
y
e
r
s
u
s
e
d
f
o
r
t
he
d
e
t
e
r
m
i
na
t
i
o
n i
n
t
he
o
ut
p
ut
l
a
ye
r
.
T
he
f
ul
l
y
-
co
n
n
ect
ed
l
a
y
er
ca
n
b
e
co
n
s
i
d
er
ed
as
t
h
e
h
i
dde
n
-
la
y
e
r
i
n
a
r
tif
ic
ia
l n
e
u
r
a
l
ne
t
w
o
r
k.
A
l
l
t
he
ne
ur
o
ns
i
n t
he
f
u
l
l
y
-
co
n
n
ect
ed
l
a
y
er
ar
e f
u
l
l
y
co
n
n
ect
ed
t
o
t
he
ne
ur
o
n
s
o
f
b
o
t
h p
r
e
vi
o
us
a
nd
ne
xt
l
a
y
e
r
s
.
T
o
r
e
duc
e
o
ve
r
f
i
t
t
i
n
g i
n t
he
f
u
l
l
y c
o
nne
c
t
e
d
l
a
ye
r
,
a
r
eg
u
l
ar
i
zat
i
o
n
m
et
h
o
d
cal
l
ed
“d
r
o
p
o
u
t
”
[
28]
i
s
u
s
u
al
l
y
e
m
p
l
o
y
ed
.
D
r
o
p
o
u
t
t
ech
n
i
q
u
e en
ab
l
es
t
h
e co
m
p
l
ex
co
-
ad
ap
t
at
i
o
n
s
o
f
n
eu
r
o
n
s
t
o
r
ed
u
ce g
r
eat
l
y
.
T
h
er
ef
o
r
e,
i
t
i
s
f
o
r
ced
t
o
l
ear
n
m
o
r
e r
o
b
u
s
t
f
eat
u
r
es
i
n
a
n
i
m
a
g
e
a
nd
e
x
h
ib
it
s
s
u
b
s
ta
n
tia
l o
v
e
r
f
itti
n
g
.
T
h
e
o
u
tp
u
t o
f
th
e
la
s
t f
u
ll
y
c
o
n
n
e
c
te
d
la
y
e
r
is
f
e
d
to
a
S
o
f
t
m
a
x
la
y
e
r
[
29]
.
T
h
is
la
y
e
r
h
a
s
5
5
o
u
tp
u
t
s
,
w
h
ic
h
p
r
o
d
u
c
e
p
r
o
b
a
b
ility
d
is
tr
ib
u
tio
n
s
c
a
lc
u
la
te
d
u
s
in
g
E
q
u
a
tio
n
2
.
T
h
e
p
r
o
b
a
b
l
i
l
i
t
y
i
s
u
s
ed
t
o
l
ab
el
each
cl
as
s
,
t
h
at
i
s
t
h
e r
eco
g
n
i
z
ed
p
er
s
o
n
.
H
en
ce,
p
r
o
b
a
b
i
l
i
t
i
es
v
ect
o
r
o
f
s
i
ze 1
×
5
5
w
h
er
e each
v
ect
o
r
e
l
e
m
e
n
t
co
r
r
es
p
o
n
d
s
t
o
a cl
as
s
o
f
d
at
as
et
i
s
o
b
t
ai
n
ed
.
=
∑
1
,
f
o
r
j
=
1
..k
(
2
)
w
h
er
e
x
i
s
t
he
ne
t
i
np
ut
.
O
ut
p
ut
va
l
ue
s
o
f
p
a
r
e
be
t
w
e
e
n
0 a
n
d 1 a
n
d t
h
e
i
r
s
um
e
qu
a
l
s
t
o 1.
3.
3
.
C
l
as
s
i
f
i
c
at
i
on
L
aye
r
T
h
i
s l
a
y
e
r
i
s
d
e
si
g
n
a
t
e
d t
o c
ou
n
t
t
h
e
e
r
r
or
(
l
os
s
)
us
i
n
g
c
r
os
s
-
e
nt
r
o
p
y
f
u
nc
t
i
o
n f
o
r
1
-
of
-
k c
odi
ng
s
ch
e
m
e as
i
n
E
q
u
a
tio
n
3.
(
)
=
−
ln
(
,
)
=
1
=
1
(
3
)
wh
e
r
e
i
s
t
h
e p
ar
a
m
e
t
er
v
ect
o
r
,
is
t
h
e
i
n
d
ic
a
to
r
th
a
t
th
e
th
s
a
m
p
le
b
e
lo
n
g
s
to
th
e
t
h
cl
as
s
,
an
d
(
,
)
is
th
e
o
u
tp
u
t
f
o
r
th
e
s
a
m
p
le
,
.
T
he
o
ut
p
ut
(
,
)
can
b
e
i
n
t
er
p
r
et
ed
as
t
h
e
p
r
o
b
ab
i
l
i
t
y
t
h
at
t
h
e
n
et
w
o
r
k
a
s
s
o
ci
at
e
s
th
i
n
p
u
t
w
ith
c
la
s
s
,
,
th
a
t i
s
,
(
=
1
|
)
.
T
he
o
ut
p
ut
u
ni
t
a
c
ti
v
a
tio
n
f
u
n
c
tio
n
i
s
t
h
e
S
o
f
t
m
a
x f
unc
t
i
o
n.
4.
RE
S
U
L
T
S
AND D
I
S
CU
S
S
I
O
NS
T
h
e
pr
o
pos
e
d C
N
N
m
ode
l
i
s
t
r
a
i
n
e
d us
i
ng
ba
t
c
h
e
s
of
i
m
a
ge
s
c
on
t
a
i
n
i
n
g
60 i
m
a
g
e
s
,
w
i
t
h
l
e
a
r
n
i
ng
r
at
e f
i
x
ed
at
0
.
0
0
1
t
o
ach
i
ev
e a r
el
i
ab
l
e t
r
ai
n
i
n
g
.
W
e u
s
ed
S
t
o
ch
as
t
i
c G
r
ad
i
en
t
D
e
s
cen
t
w
i
t
h
M
o
me
n
t
u
m
f
o
r
t
r
a
i
ni
ng t
he
C
N
N
.
T
he
va
l
ue
o
f
m
o
m
e
nt
u
m
i
s
0
.
9
i
n
s
p
i
r
e
d
b
y
[
29]
-
[3
1
].
T
h
e
m
a
x
i
m
u
m
n
um
be
r
of
e
poc
h
i
s
s
e
t
t
o
5
0
t
o
e
ns
ur
e
ne
t
w
o
r
k
c
o
nve
r
ge
nc
e
.
T
he
t
r
a
i
ni
ng
a
nd
va
l
i
d
a
t
i
o
n
o
ut
c
o
m
e
s
a
r
e
p
r
e
s
e
nt
e
d
i
n
F
i
g
ur
e
6
.
S
i
n
ce t
h
e
C
N
N
m
o
d
el
i
s
t
r
ai
n
ed
f
r
o
m
s
cr
at
ch
,
t
h
e r
eco
g
n
i
t
i
o
n
accu
r
ac
y
s
t
ar
t
ed
w
i
t
h
a
v
al
u
e c
l
os
e
t
o 0 a
n
d
g
r
ad
u
al
l
y
i
m
p
r
o
v
es
al
o
n
g
w
i
t
h
t
h
e
n
u
m
b
er
o
f
i
t
er
at
i
o
n
s
.
A
f
t
er
i
t
er
at
i
o
n
5
,
0
0
0
,
t
h
e acc
u
r
ac
y
s
t
ab
i
l
i
zed
at
c
l
o
s
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
25
02
-
4752
I
nd
o
ne
s
i
a
n J
E
l
e
c
E
ng
&
C
o
m
p
S
c
i
,
V
o
l.
11
, N
o
.
2
,
A
ug
us
t
2018
:
5
58
–
566
564
t
o
v
al
u
e 1
0
0
%
.
T
h
er
ef
o
r
e,
t
h
e t
r
ai
n
i
n
g
i
s
s
t
o
p
p
ed
at
v
al
i
d
at
i
o
n
acc
u
r
ac
y
o
f
9
9
.
7
6
%
as
f
u
r
t
h
er
t
r
ai
n
i
n
g
i
s
d
eem
ed
u
n
n
ece
s
s
ar
y
.
T
h
e
lo
s
s
r
a
te
is
i
llu
s
tr
a
te
d
i
n
F
i
g
u
r
e
7
.
A
t
t
h
e ear
l
y
s
t
a
g
e o
f
t
r
ai
n
i
n
g
,
t
h
e l
o
s
s
r
at
e i
s
r
at
h
er
h
i
g
h
at
a v
al
u
e o
f
cl
o
s
e
t
o
6
0
.
A
s
t
h
e t
r
ai
n
i
n
g
p
r
o
ceed
ed
,
t
h
e l
o
s
s
r
at
e g
r
ad
u
al
l
y
d
ecr
eas
es
ev
en
t
u
al
l
y
r
each
i
n
g
t
o
l
o
s
s
v
al
u
e cl
o
s
e t
o
0
.
F
i
g
ur
e
6
.
T
ra
i
n
i
ng
a
n
d v
a
l
i
da
t
i
on
of
t
h
e
pr
opos
e
d C
N
N
m
ode
l
F
i
g
ur
e
7
.
L
os
s
r
a
t
e
of
t
h
e
pr
op
os
e
d C
N
N
m
ode
l
F
i
g
ur
e
8.
W
r
o
ngl
y
r
e
c
o
g
ni
z
e
d
s
ub
j
e
c
t
s
w
i
t
h i
t
s
c
o
r
r
e
s
p
o
nd
i
n
g l
u
m
i
na
nc
e
va
l
ue
I
dent
i
f
i
ed
as
:
#31
A
c
t
ual
:
#6
Lum
i
nan
c
e:
6
l
ux
I
dent
i
f
i
ed as
:
#48
A
c
t
ual
:
#29
Lum
i
nan
c
e:
5
l
ux
I
dent
i
f
i
ed as
:
#38
A
c
t
ual
:
#14
Lum
i
nan
c
e:
22 l
ux
I
dent
i
f
i
ed as
:
#4
A
c
t
ual
:
#12
Lum
i
nan
c
e:
3
l
ux
I
dent
i
f
i
ed
as
:
#11
A
c
t
ual
:
#30
Lum
i
nan
c
e:
2
l
ux
I
dent
i
f
i
ed
as
:
#28
A
c
t
ual
:
#38
Lum
i
nan
c
e:
3
l
ux
I
dent
i
f
i
ed as
:
#41
A
c
t
ual
:
#38
Lum
i
nan
c
e:
3
l
ux
I
dent
i
f
i
ed as
:
#4
A
c
t
ual
:
#45
Lum
i
nan
c
e:
10 l
ux
A
ccu
r
acy
s
t
ab
i
l
i
zes
a
f
t
er
i
t
e
r
a
t
i
o
n
5
0
00
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nd
o
ne
s
i
a
n J
E
l
e
c
E
ng
&
C
o
m
p
S
c
i
I
SSN
:
2502
-
4752
C
an C
onv
ol
ut
i
on N
e
ur
al
N
e
t
w
or
k
(
C
N
N
)
T
r
i
um
ph i
n E
ar
R
e
c
ogni
t
i
on
of
U
ni
f
or
m
…
(
Nu
r
s
u
ri
a
ti J
a
m
il
)
565
F
o
r
t
es
t
i
n
g
,
a t
o
t
al
o
f
2
,
7
5
0
ear
i
m
a
g
es
ar
e
u
s
ed
f
o
r
r
eco
g
n
i
t
i
o
n
a
n
d
t
h
e r
es
ul
t
s
ho
w
s
t
ha
t
2
,
7
4
2
i
m
a
g
es
ar
e
co
r
r
ect
l
y
r
eco
g
n
i
ze
d
.
F
i
g
u
r
e 8
i
l
l
u
s
t
r
at
es
t
h
e
w
r
o
n
g
l
y
r
eco
g
n
i
zed
8
s
u
b
j
ect
s
w
i
t
h
i
t
s
co
r
r
es
p
o
n
d
i
n
g
l
u
m
i
n
a
n
ce
v
al
u
e.
A
s
ca
n
b
e s
een
,
w
i
t
h
t
h
e e
x
cep
t
i
o
n
o
f
o
n
e i
m
a
g
e,
al
l
t
h
e
w
r
o
n
g
l
y
r
eco
g
n
i
zed
i
m
a
g
e
s
h
av
e
l
ux
va
l
ue
s
o
f
l
e
s
s
t
h
a
n
1
0
.
I
n
t
o
t
al
,
t
h
er
e ar
e 2
3
6
ear
i
m
a
g
e
s
(
R
e
f
er
T
ab
l
e 1
)
w
i
t
h
t
h
e l
u
x
v
al
u
es
o
f
l
es
s
t
h
a
n
10,
a
n
d
ou
r
pr
opos
e
d
C
N
N
m
ode
l
i
s
a
bl
e
t
o r
e
c
og
n
i
z
e
229
i
m
a
g
e
s
(
97%
)
c
or
r
e
c
t
l
y
.
E
v
e
n t
h
ou
gh
t
h
e
e
a
r
i
n
t
he
i
m
a
ge
i
s
no
t
e
ve
n
vi
s
i
b
l
e
b
y na
ke
d
e
ye
s
,
t
he
p
r
o
p
o
s
e
d
m
o
d
el
i
s
ab
l
e t
o
co
r
r
ect
l
y
r
eco
g
n
i
zed
m
aj
o
r
i
t
y
o
f
t
h
e
s
u
bj
e
c
t
s
.
A
s
f
or
s
u
bj
e
c
t
#38,
t
h
e
a
u
gm
e
n
t
e
d l
e
f
t
e
a
r
of
s
u
bj
e
c
t
#14 i
s
w
r
o
n
gl
y
r
e
c
og
n
i
z
e
d
e
v
e
n
t
h
o
u
gh
a
de
qu
a
t
e
l
um
i
n
a
n
c
e
of
22 l
u
x
i
s
r
e
c
or
de
d
f
or
t
h
e
i
m
a
g
e
.
T
h
i
s
i
s
pr
oba
bl
y
du
e
t
o t
h
e
r
ot
a
t
e
d
ope
r
a
t
i
on
p
er
f
o
r
m
ed
o
n
t
h
e i
m
ag
e.
5.
CO
NCL
U
S
I
O
N
D
e
e
p
l
e
a
r
ni
ng
u
s
i
n
g
c
o
nvo
l
ut
i
o
n
ne
ur
a
l
ne
t
w
o
r
k
(
C
N
N
)
ha
s
s
ho
w
n
t
o
p
r
o
d
uc
e
e
xc
e
p
t
i
o
na
l
pe
r
f
or
m
a
n
c
e
ov
e
r
t
r
a
di
t
i
on
a
l
m
e
t
h
ods
i
n
m
a
ny
a
ppl
i
c
a
t
i
on
dom
a
i
n
s
[
32]
.
I
n
t
h
i
s
pa
pe
r
,
a
C
N
N
m
ode
l
i
s
bui
l
t
an
d
t
r
ai
n
ed
f
o
r
ear
b
i
o
m
e
t
r
i
c
s
i
n v
a
r
i
ous
un
i
f
or
m
i
l
l
um
i
n
a
t
i
on
s
m
e
a
s
u
r
e
d us
i
n
g l
um
e
ns
.
A
s
f
a
r
a
s
w
e
k
n
o
w
,
th
is
is
th
e
f
ir
s
t
w
o
r
k
o
f
te
s
t
in
g
t
h
e
p
e
r
f
o
r
m
a
n
c
e
o
f
C
N
N
o
n
v
e
r
y
un
de
r
e
x
pos
e
d or
ov
e
r
e
x
pos
e
d i
m
a
g
e
s
.
O
ur
r
es
u
l
t
s
s
h
o
w
ed
t
h
at
f
o
r
i
m
ag
es
u
n
i
f
o
r
m
l
y
i
l
l
u
m
i
n
at
ed
w
i
t
h
l
u
m
i
n
an
ce o
f
a
bov
e
25 l
u
x,
10
0%
r
e
c
og
n
i
t
i
on
a
c
c
u
r
a
c
y
i
s
a
c
h
i
e
v
e
d.
E
v
e
n
t
hou
gh
t
h
e
C
N
N
m
ode
l
h
a
s
pr
obl
e
m
s
r
e
c
ogn
i
z
i
n
g
f
e
w
i
m
a
g
e
s
of
be
l
o
w
10 l
ux
,
t
h
e o
v
er
al
l
accu
r
ac
y
o
f
9
7
%
s
u
g
g
es
t
s
t
h
a
t
C
N
N
ar
ch
i
t
ect
u
r
e
p
er
f
o
r
m
s
eq
u
al
l
y
w
el
l
f
o
r
u
n
i
f
o
r
m
i
l
l
u
m
i
n
at
i
o
n
-
i
nva
r
i
a
nt
i
m
a
ge
s
.
H
o
w
e
v
e
r
,
da
t
a
a
u
gm
e
n
t
a
t
i
o
n
f
or
ou
r
da
t
a
s
e
t
on
l
y
i
nv
ol
v
e
s
r
ot
a
t
i
o
n
s
.
M
or
e
r
obu
s
t
ope
r
a
t
i
on
s
s
h
oul
d be
i
n
c
l
u
de
d i
n
da
t
a
a
ugm
e
nt
a
t
i
on
t
o ul
t
i
m
a
t
e
l
y
t
e
s
t
t
he
pe
r
f
or
m
a
n
c
e
of
C
N
N
m
ode
l
f
or
e
a
r
bi
om
e
t
r
i
c
s
.
ACK
NO
W
L
E
D
G
E
M
E
NT
S
D
u
e ac
k
n
o
w
l
ed
g
m
en
t
i
s
acco
r
d
ed
t
o
t
h
e
R
es
ear
c
h
Ma
n
ag
e
m
en
t
C
en
t
er
an
d
F
acu
l
t
y
o
f
C
o
m
p
u
t
er
&
M
at
h
e
m
at
i
cal
S
ci
e
n
ce
s
,
U
n
i
v
e
r
s
i
t
i
T
ek
n
o
l
o
g
i
M
A
R
A
,
f
o
r
t
h
e f
u
n
d
i
n
g
r
ecei
v
ed
t
h
r
o
u
g
h
t
h
e B
E
S
T
A
R
I
G
r
an
t
,
(
600
-
I
R
M
I
/
D
AN
A 5
/
3/
B
E
S
T
A
R
I
(
0002/
2
016
)
a
n
d
U
n
iv
e
r
s
titi T
e
n
a
g
a
N
a
s
io
n
a
l.
R
EF
ER
EN
C
ES
[
1]
Hu
b
e
l
DH,
W
i
e
s
e
l
T
N
.
R
e
c
e
p
tiv
e
F
ie
ld
s
o
f
S
in
g
le
N
e
u
r
o
n
e
s
in
th
e
C
a
t’
s
S
tr
ia
te
C
o
r
te
x
.
T
he
J
our
nal
of
.
P
hy
s
i
ol
ogy
.
195
9;
14
8:
57
4
-
59
1.
[
2]
F
uk
us
hi
m
a
K
.
N
e
oc
og
ni
t
r
o
n:
A
S
e
l
f
-
O
r
g
an
i
zi
n
g
N
eu
r
al
N
et
w
o
r
k
M
o
d
el
f
o
r
a M
ech
an
i
s
m
o
f
P
at
t
er
n
R
eco
g
n
i
t
i
o
n
U
na
f
f
e
c
t
e
d by
S
hi
f
t
i
n
P
os
i
t
i
o
n
,
”
B
i
ol
.
C
y
be
r
n.
19
80;
3
6(
4)
:
1
93
–
2
02.
[
3]
L
e
C
un Y
,
B
e
ng
i
o Y
.
C
onv
ol
ut
i
o
na
l
ne
t
w
or
k
s
f
or
i
m
a
ge
s
,
s
pe
e
c
h,
a
nd t
i
m
e
s
e
r
i
e
s
.
T
h
e
H
and
bo
ok
of
B
r
ai
n T
he
or
y
and
N
e
ur
al
N
e
t
w
or
k
s
.
1
99
5 A
pr
;
336
1(
10)
:
19
95.
[
4]
L
e
C
un Y
,
B
ot
t
o
u L
,
B
e
ng
i
o Y
,
H
a
f
f
ne
r
P
.
Gra
d
i
e
n
t
-
B
as
e
d L
e
ar
ni
n
g
A
p
pl
i
e
d t
o D
oc
um
e
nt
R
e
c
o
gn
i
t
i
on
.
P
r
o
ceed
i
n
g
s
of
t
he
I
E
E
E
.
19
98
N
ov
;
8
6(
1
1)
:
2
278
-
3
24.
[
5]
L
e
C
un Y
,
H
ua
ng
F
J
,
B
ot
t
o
u L
.
L
e
ar
ni
n
g M
e
t
ho
ds
f
or
G
e
ne
r
i
c
O
bj
e
c
t
R
e
c
ogni
t
i
o
n w
i
t
h I
nv
ar
i
a
nc
e
t
o P
os
e
a
n
d
L
ig
h
tin
g
.
P
r
oc
e
e
di
ng
s
of
t
he
20
04
I
E
E
E
C
om
put
e
r
V
i
s
i
on
a
nd
P
a
t
t
e
r
n R
e
c
og
ni
t
i
on,
2
00
4;
2:
I
I
-
1
04.
[
6]
H
ech
t
-
N
i
e
l
s
e
n R
.
T
he
or
y
of
t
he
B
a
c
k
pr
opa
g
a
t
i
on N
e
ur
a
l
N
e
t
w
or
k
.
I
n N
e
ur
a
l
N
e
t
w
or
k
s
f
or
P
e
r
c
e
pt
i
o
n;
1
99
2:
65
-
9
3.
[
7]
S
un
M
,
S
ong
Z
,
J
i
a
ng
X
,
P
a
n J
,
P
a
ng
Y
.
L
e
a
r
ni
ng
P
o
ol
i
ng
f
or
C
onv
ol
ut
i
o
na
l
N
e
ur
a
l
N
e
t
w
or
k
.
N
e
ur
oc
om
p
ut
i
n
g
.
201
7 F
e
b
8;
22
4:
96
-
10
4.
[
8]
S
ch
er
er
D
,
M
ü
l
l
er
A
,
B
eh
n
k
e
S
.
E
v
al
uat
i
o
n of
P
o
ol
i
ng O
pe
r
a
t
i
ons
i
n C
onv
ol
ut
i
on
al
A
r
c
hi
t
e
c
t
ur
e
s
f
or
O
bj
e
c
t
R
e
c
ogni
t
i
o
n
.
I
nt
e
r
na
t
i
ona
l
C
o
nf
e
r
e
nc
e
on
A
r
t
i
f
i
c
i
a
l
N
e
ur
a
l
N
e
t
w
or
k
s
,
S
pr
i
ng
e
r
,
B
e
r
l
i
n,
H
e
i
de
l
be
r
g
;
2010 S
e
p 1
5:
92
-
10
1.
[
9]
K
r
i
zh
ev
s
k
y
A
,
S
u
ts
k
e
v
e
r
I
,
H
in
to
n
G
E.
I
m
a
g
e
n
e
t C
la
s
s
i
f
ic
a
tio
n
w
ith
D
e
e
p
C
o
n
v
o
lu
ti
o
n
a
l
N
e
u
r
a
l N
e
tw
o
r
k
s
.
I
n
A
dv
a
nc
e
s
i
n N
e
ur
a
l
I
nf
or
m
a
t
i
on
P
r
oc
e
s
s
i
ng
S
y
s
t
e
m
s
;
201
2:
1
09
7
-
110
5.
[
1
0]
L
ee S
H
,
C
h
an
C
S
,
M
a
y
o
S
J
,
R
e
m
ag
n
i
n
o
P
.
H
o
w
D
eep
L
ear
n
i
n
g
E
x
t
r
act
s
an
d
L
e
ar
n
s
L
e
a
f
F
e
at
u
r
es
f
or
P
l
a
nt
C
la
s
s
if
ic
a
tio
n
.
P
a
tte
r
n
Re
c
o
g
n
itio
n
.
2
01
7 N
ov
30;
7
1:
1
-
3.
[
1
1]
B
ouf
e
na
r
C
,
K
e
r
boua
A
,
B
a
t
ouc
he
M
.
I
nv
e
s
t
i
g
a
t
i
on on D
e
e
p L
e
a
r
ni
ng
f
or
O
f
f
-
l
i
n
e H
an
d
w
r
i
t
t
en
A
r
ab
i
c C
h
ar
act
er
R
e
c
o
g
n
itio
n
.
C
o
g
n
i
t
i
ve S
ys
t
em
s
R
es
ea
r
ch
.
20
17 N
ov
1
3.
[
1
2]
B
a
k
r
y
A
, E
l
h
o
s
e
i
n
y
M
, E
l
-
G
a
al
y
T
,
E
l
g
a
m
m
al
A
.
D
i
ggi
n
g D
e
e
p i
nt
o t
he
L
ay
e
r
s
of
C
N
N
s
:
I
n Se
ar
c
h of
H
ow
C
N
N
A
ch
i
eve V
i
ew
I
n
va
r
i
a
n
ce
.
P
r
oc
e
e
di
ng
s
of
t
he
I
nt
e
r
na
t
i
ona
l
C
onf
e
r
e
nc
e
on L
e
a
r
ni
ng
R
e
pr
e
s
e
nt
a
t
i
on
s
(
I
C
L
R
2016)
a
r
X
i
v
pr
e
pr
i
nt
a
r
X
i
v
:
1
50
8.
0
19
83
.
C
or
ne
l
l
U
ni
v
e
r
s
i
t
y
L
i
br
a
r
y
.
201
5
A
u
g 9
.
[
1
3]
L
a
i
K
, B
o
L
, R
e
n
X
, F
o
x
D
.
A
L
ar
ge
-
S
c
a
le
H
ie
r
a
r
c
h
ic
a
l M
u
lti
-
vi
ew
R
G
B
-
D Ob
j
e
c
t
Da
t
a
se
t
.
P
r
oc
e
e
di
ng
s
of
t
he
201
1 I
E
E
E
I
nt
e
r
na
t
i
o
na
l
C
onf
e
r
e
nc
e
on R
ob
ot
i
c
s
a
n
d A
ut
om
a
t
i
on
(
I
C
R
A
)
,
2011
M
a
y
9;
18
17
-
18
24.
[
1
4]
X
i
an
g
Y
,
M
o
t
t
ag
h
i
R
,
S
av
ar
es
e
S
.
B
e
y
o
n
d
P
a
s
c
a
l
:
A
B
e
n
c
h
m
a
r
k
fo
r
3
D
O
b
je
c
t D
e
te
c
tio
n
in
th
e
W
ild
.
2014
P
r
oc
e
e
di
ng
s
of
t
he
I
E
E
E
W
i
nt
e
r
C
onf
e
r
e
nc
e
on A
ppl
i
c
a
t
i
o
ns
of
C
om
put
e
r
V
i
s
i
on
(
W
A
C
V
)
,
201
4
M
a
r
24;
7
5
-
82
.
[
1
5]
B
o
r
ji A
,
I
z
a
d
i S
,
I
tti L
.
W
hat
C
a
n W
e
L
e
ar
n A
b
out
C
N
N
s
F
r
om
a L
ar
ge
Sc
al
e
C
o
nt
r
o
l
l
e
d O
bj
e
c
t
D
at
as
e
t
?
a
rX
i
v
pr
e
pr
i
n
t
a
r
X
i
v
:
15
12.
01
32
0.
C
or
n
e
l
l
U
ni
v
e
r
s
i
t
y
L
i
br
a
r
y
.
2015
D
e
c
4.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
25
02
-
4752
I
nd
o
ne
s
i
a
n J
E
l
e
c
E
ng
&
C
o
m
p
S
c
i
,
V
o
l.
11
, N
o
.
2
,
A
ug
us
t
2018
:
5
58
–
566
566
[
1
6]
M
a
dde
r
n W
,
S
t
e
w
a
r
t
A
,
M
c
M
a
nus
C
,
U
pc
r
of
t
B
,
C
hur
c
hi
l
l
W
,
N
e
w
m
a
n P
.
I
l
l
um
i
nat
i
on
I
n
v
ar
i
ant
I
m
agi
ng:
A
ppl
i
c
at
i
ons
i
n R
ob
us
t
V
i
s
i
on
-
B
as
e
d L
oc
al
i
s
at
i
o
n,
M
ap
pi
n
g a
nd C
l
as
s
i
f
i
c
at
i
on f
or
A
ut
on
om
o
us
V
e
hi
c
l
e
s
.
I
n
P
r
oc
e
e
di
ng
s
of
t
he
I
E
E
E
I
nt
e
r
n
a
t
i
ona
l
C
o
nf
e
r
e
nc
e
on R
obo
t
i
c
s
a
nd A
ut
om
a
t
i
on V
i
s
ua
l
P
l
a
c
e
R
e
c
og
ni
t
i
on i
n
C
ha
ng
i
ng
E
nv
i
r
o
nm
e
nt
s
W
or
k
s
hop,
(
I
C
R
A
)
.
H
ong
K
ong
,
C
hi
na
.
20
1
4,
M
a
y
;
2
:
3.
[
1
7]
Y
o
n
g
L
,
Y
u
an
zh
i
C
.
A
L
o
cal
F
eat
u
r
e D
es
cr
i
p
t
o
r
I
n
v
ar
i
an
t
t
o C
om
pl
e
x
I
l
l
u
m
i
na
t
i
on C
ha
ng
e
s
.
T
el
ko
m
n
i
ka
(
T
e
l
e
c
om
m
uni
c
a
t
i
on C
om
put
i
ng
E
l
e
c
t
r
oni
c
s
a
n
d C
o
nt
r
o
l
)
.
20
16 J
u
n 2;
1
4(
2A
)
:
28
9
-
97
.
[
1
8]
V
o
ni
k
a
k
i
s
V
,
C
hr
y
s
os
t
om
ou D
,
K
ous
k
our
i
da
s
R
,
G
a
s
t
e
r
a
t
os
A
.
A
B
i
ol
og
i
c
a
l
l
y
I
ns
pi
r
e
d S
c
a
l
e
-
S
p
a
c
e
f
o
r
I
llu
m
in
a
tio
n
I
n
v
ar
i
an
t
F
eat
u
r
e
D
et
ect
i
o
n
.
M
ea
s
u
r
em
en
t
S
ci
en
ce a
n
d
T
ech
n
o
l
o
g
y
.
20
13
J
u
n 1
2;
24(
7)
:
0
74
02
4.
[
1
9]
E
m
er
š
i
č
Ž
,
Š
t
ep
ec
D
,
Š
t
r
u
c
V
,
P
e
er
P
.
T
r
ai
n
i
n
g C
onv
ol
ut
i
on
al
N
e
u
r
al
N
e
t
w
or
k
s
w
i
t
h
L
i
m
i
t
e
d
T
r
ai
ni
ng D
at
a
f
or
E
ar
Re
c
o
g
n
it
io
n
in
th
e
W
il
d
.
a
r
X
i
v
pr
e
pr
i
nt
a
r
X
i
v
:
1
71
1.
0
99
52
.
C
or
ne
l
l
U
ni
v
e
r
s
i
t
y
L
i
br
a
r
y
.
2017 N
ov
27
.
[
2
0]
E
m
e
r
š
i
č
Ž
, Š
t
r
u
c
V
,
P
e
e
r
P
. E
a
r
R
e
c
o
g
n
i
t
i
o
n
:
M
o
r
e
T
h
a
n
a
S
u
r
v
e
y
.
N
e
ur
oc
om
p
ut
i
n
g
.
20
17
S
e
p 13;
2
55:
26
-
3
9.
[
2
1]
Z
ha
ng
Y
,
M
u Z
.
E
a
r
D
e
t
e
c
t
i
on U
nde
r
U
nc
ont
r
ol
l
e
d C
on
di
t
i
ons
w
i
t
h M
ul
t
i
pl
e
S
c
a
l
e
F
a
s
t
e
r
R
e
g
i
on
-
B
as
ed
C
onv
ol
ut
i
o
na
l
N
e
ur
a
l
Ne
t
w
o
r
k
s.
S
ym
m
et
r
y.
201
7 A
pr
1
0;
9
(
4)
:
53
.
[
2
2]
G
al
d
á
m
ez P
L
,
R
a
v
ean
e
W
,
A
r
r
i
et
a
A
G
.
A
B
r
i
e
f
R
e
v
i
e
w
o
f
t
h
e E
ar
R
e
co
g
n
i
t
i
o
n
P
r
o
ces
s
u
s
i
n
g
D
e
ep
N
eu
r
al
Ne
t
w
o
r
k
s.
J
our
na
l
of
A
p
pl
i
e
d
L
o
gi
c
.
2
01
7
N
ov
1;
24:
62
-
7
0.
[
2
3]
D
e
v
lin
K
,
C
h
a
l
m
e
r
s
A
,
R
e
in
h
a
r
d
E.
V
is
u
a
l C
a
lib
r
a
tio
n
a
nd C
or
r
e
c
t
i
on f
or
A
m
bi
e
nt
I
l
l
um
i
na
t
i
on.
A
C
M
T
r
ans
ac
t
i
o
ns
on A
pp
l
i
e
d P
e
r
c
e
pt
i
on (
T
A
P
)
.
2
0
06 O
c
t
1;
3(
4)
:
4
29
-
4
52
.
[
2
4]
A
r
i
f
f
i
n
S
M
,
Ja
m
i
l
N,
R
a
h
m
a
n
P
N.
D
I
A
ST
V
ar
i
abi
l
i
t
y
I
l
l
um
i
nat
e
d T
he
r
m
al
a
nd V
i
s
i
bl
e
E
ar
I
m
a
ge
s
D
at
as
e
t
s
.
I
n
P
r
o
ceed
i
n
g
s
o
f
t
h
e I
E
E
E
S
i
g
n
al
P
r
o
ces
s
in
g
: A
lg
o
r
ith
m
s
,
A
r
c
h
ite
c
tu
r
e
s
,
A
r
r
a
n
g
e
m
e
n
ts
,
a
n
d
A
p
p
lic
a
tio
n
s
(
S
P
A
)
,
201
6 S
e
p
2
1;
19
1
-
19
5.
[
2
5]
K
r
i
z
he
v
s
ky
A
,
S
ut
s
k
e
v
e
r
I
,
H
i
nt
on G
E
.
I
m
a
g
e
n
e
t C
la
s
s
ific
a
tio
n
w
ith
D
e
e
p
C
o
n
v
o
l
u
tio
n
a
l N
e
u
r
a
l N
e
tw
o
r
k
s
.
In
A
dv
a
nc
e
s
i
n N
e
ur
a
l
I
nf
or
m
a
t
i
on
P
r
oc
e
s
s
i
ng
S
y
s
t
e
m
s
2012
;
1
0
97
-
11
0
5.
[
2
6]
I
o
f
f
e S
,
S
ze
g
ed
y
C
.
B
at
c
h N
or
m
al
i
z
at
i
on:
A
c
c
e
l
e
r
at
i
n
g D
e
e
p N
e
t
w
or
k
T
r
ai
ni
n
g by
R
e
duc
i
n
g I
nt
e
r
nal
C
ov
ar
i
at
e
S
h
ift
.
I
n
P
r
oc
e
e
di
ng
s
of
I
nt
e
r
na
t
i
o
na
l
c
onf
e
r
e
nc
e
on M
a
c
h
i
ne
L
e
a
r
ni
ng
,
2
01
5 J
u
n 1
;
4
48
-
45
6.
[
2
7]
N
a
i
r
V
, H
i
n
t
o
n
G
E
.
Re
c
tifie
d
L
i
n
e
a
r
U
n
its
I
m
p
r
o
ve R
es
t
r
i
ct
ed
B
o
l
t
z
m
a
n
n
M
a
ch
i
n
es
.
I
n
P
r
oc
e
e
di
ng
s
of
t
he
27t
h
i
nt
e
r
na
t
i
ona
l
C
onf
e
r
e
nc
e
on M
a
c
hi
ne
L
e
a
r
ni
ng
(
I
C
M
L
-
10)
,
2
01
0;
807
-
81
4.
[
2
8]
S
r
i
v
a
s
t
a
v
a
N
,
H
i
nt
on
G
,
K
r
i
z
he
vs
ky
A
,
S
ut
s
k
e
ve
r
I
,
S
a
l
a
k
hut
di
n
o
v
R
.
D
r
opo
ut
:
A
S
i
m
pl
e
W
a
y
t
o P
r
e
v
e
nt
N
e
ur
a
l
Ne
t
w
o
r
k
s f
r
o
m
O
v
e
r
f
ittin
g
.
T
he
J
our
n
al
of
M
ac
hi
ne
L
e
ar
ni
n
g R
e
s
e
ar
c
h
.
20
14
J
a
n
1;
15(
1)
:
19
29
-
5
8.
[
2
9]
C
h
r
i
s
t
o
p
h
er
M
B
.
P
at
t
er
n
R
eco
g
n
i
t
i
o
n
an
d
M
ach
i
n
e L
ear
n
i
n
g
.
S
p
r
i
n
g
er
-
V
e
r
l
a
g
N
e
w
Y
or
k
;
2016
.
[
3
0]
G
u J
,
W
a
ng
Z
,
K
ue
n J
,
M
a
L
,
S
ha
hr
ou
dy
A
,
S
hua
i
B
,
L
i
u T
,
W
a
ng
X
,
W
a
ng
G
,
C
a
i
J
,
C
he
n T
.
R
e
c
e
nt
A
dv
a
nc
e
s
i
n
C
onv
ol
ut
i
o
na
l
N
e
ur
a
l
N
e
t
w
or
k
s
.
Pa
tte
r
n
Re
c
o
g
n
itio
n
.
2
01
7 O
c
t
11.
[
3
1]
D
e
ng
W
,
F
a
ng
Y
,
X
u Z
,
H
u J
.
F
a
c
i
a
l
L
a
ndm
a
r
k
L
oc
a
l
i
z
a
t
i
on by
E
nha
nc
e
d C
onv
ol
u
t
i
o
na
l
N
eu
r
al
N
et
w
o
r
k
.
N
e
ur
oc
om
p
ut
i
n
g
.
20
18
J
a
n 1
7;
27
3:
2
22
-
9.
[
3
2]
B
or
od
o S
M
,
S
ha
m
s
uddi
n S
M
,
H
a
s
a
n S
.
B
i
g
D
a
t
a
P
l
a
t
f
or
m
s
a
nd T
e
c
hni
q
ue
s
.
I
nd
one
s
i
a
n J
our
nal
o
f
E
l
e
c
t
r
i
c
al
E
ngi
ne
e
r
i
ng
an
d C
om
p
ut
e
r
Sc
i
e
n
c
e
.
201
6 J
a
n
1;
1(
1)
:
1
91
-
20
0.
Evaluation Warning : The document was created with Spire.PDF for Python.